Optimality of empirical measures as quantizers

Abstract: A common way to discretize a probability measure is to use an empirical measure as a discrete approximation. But how far from being optimal is this approximation in the p-Wasserstein distance? In this paper, we study this question in two contexts: (1) optimality among all uniform quantizers and (2) optimality among all (non-uniform) quantizers. In the first context, for p=1, we provide a complete answer to this question up to a polylog(n) factor. From the probabilistic point of view, this resolves, up to a polylog(n) factor, the problem of characterizing the expected 1-Wasserstein distance between a probability measure and its empirical measure in terms of non-random quantities. We also obtain some partial results for p>1 in the first context and for p>=1 in the second context.